AI wrote my compiler. A mathematical proof checks its work on every build.

Short summary

The author built Almide, a programming language designed for AI agents to write correctly, and used Claude models to develop its compiler over 4.5 months. The key metric is Modification Survival Rate: whether code converges to working state within 3 compile-retry cycles, not first-try accuracy. Empirical grammar experiments (e.g., fn-lambda vs paren-lambda) showed no meaningful difference between syntaxes, confirming that frontier models rarely fail on syntax — failures are semantic.

• •Almide is a language designed for AI agents to write, with its compiler largely co-authored by Claude models
• •Modification Survival Rate measures retry-loop convergence, not first-draft correctness — the project's highest-leverage decision
• •Empirical A/B testing of syntax choices via grammar-lab showed frontier models fail on semantics, not syntax

Generated with AI, which can make mistakes.